Skip to main content

MCP Dev Summit NA 2026 — Conference Brief

Executive Summary​

The MCP Dev Summit North America 2026 (April 1-3, New York City) marked a turning point for the Model Context Protocol ecosystem. Across three days of keynotes, technical sessions, and workshops, a clear consensus emerged: MCP has crossed the threshold from experimental protocol to enterprise production infrastructure. With 100M+ SDK downloads per month, major deployments at Amazon, Uber, and Duolingo, and governance now under the Agentic AI Foundation, the conference painted a picture of a protocol ecosystem that is maturing rapidly. The dominant architectural pattern — a gateway + registry control plane — appeared independently across nearly every enterprise talk, suggesting convergence on how organizations will manage AI agent integrations at scale.

Key Themes and Important Ideas​

1. MCP Growth and Ecosystem Maturity​

David Soria Parra, co-creator of MCP at Anthropic, opened with headline metrics: 100M+ SDK downloads per month across Python, TypeScript, Java, and other language SDKs. MCP achieved in 16 months what took React approximately 3 years — a signal that the ecosystem need for a standard integration layer between AI agents and tools is genuine and urgent. The protocol has evolved significantly since launch, adding remote server support, authorization frameworks, structured outputs, long-running tasks, SDK stability tiers, and an extensions mechanism for safe experimentation.

"MCP reached in 16 months what took React approximately 3 years to achieve." — David Soria Parra, Anthropic

2. Enterprise as the Real Deployment Surface​

A recurring theme was that the most significant MCP deployments are invisible. Behind corporate firewalls, teams are wiring MCP servers to CRMs, wikis, Snowflake warehouses, and internal knowledge bases. These internal deployments represent some of the largest MCP usage — far exceeding what is visible on social media or open-source registries. Amazon's James Hood noted that MCP is "the most popular way to connect agents to internal systems" among tens of thousands of internal builders. Uber reported 5,000+ engineers with 90%+ using AI weekly, 1,500+ internal agents, and 60,000+ agent executions per week.

"MCP is the most popular way to connect agents to internal systems." — James Hood, AWS

3. Gateway + Registry as Essential Architecture​

The gateway + registry pattern emerged independently in talks from AWS, Uber, and Obot — a strong signal of architectural convergence. The pattern consists of two components:

  • MCP Gateway: A control point that mediates all MCP interactions, providing access control, PII redaction, authorization integration, and full observability.
  • MCP Registry: A centralized catalog for discovering, sharing, and managing MCP servers and agent configurations.

Uber built a gateway that automatically translates service endpoints to MCP tools while keeping service owners in control. Amazon built an internal registry that treats agent configurations as first-class primitives, enabling security scanning at the configuration level. Sheng Liang (Obot) framed the gateway and registry as the two essential pieces of technology every enterprise needs.

"Every enterprise needs two things: a gateway to control what's happening, and a registry as a single source of truth." — Sheng Liang, Obot

4. Agent Control Plane​

Workato's Adam Seligman (CTO) and Zayne Turner introduced the Agent Control Plane — a framework for balancing the autonomy of probabilistic AI reasoning with the governance requirements of deterministic enterprise systems. The framework separates concerns into two layers: the reasoning layer (LLM, agents) which is probabilistic, and the control plane which is deterministic and handles governance, authentication, and business logic.

Their Seven Factors for Agent Control Plane provide a practical checklist: Governance Operations, Deterministic Mutations, Intent-Based Communication, Founded Access, Safe Retries, Appropriate Errors, and Structural Observability. The white paper is available at workato.com/sevenfactors.

"Protocols are thin. Everything the enterprise needs — governance, auth, business logic — you have to handle yourself." — Adam Seligman, Workato

5. Context Window is Solvable​

A question that came up repeatedly — "how do you handle hundreds of tools in a single context window?" — was answered consistently: this is a client implementation problem, not a protocol problem. Amazon's approach uses progressive disclosure: inject critical tools directly into the context, wrap others in CLI-style interfaces for on-demand discovery. Duolingo connects 180+ tools to a single Slackbot through a centralized "MCP store" that handles discovery. The protocol does not need to change; clients need to get smarter about what they surface.

6. Security Architecture​

Security was woven through every enterprise talk. Amazon highlighted Simon Willison's concept of the "lethal trifecta" — agent configurations that combine private data access, untrusted content ingestion, and external communication capabilities. When all three are present, the risk of data exfiltration is high. Amazon's registry scans configurations for these properties.

Uber integrates PII redaction and authorization directly into their MCP gateway. Duolingo enforces security principles: no write operations without human approval, sandboxed environments, and no side-channel communication between users. Obot outlined five security layers: MCP gateway, LLM gateway, software supply chain filtering, isolated runtime, and elevated access control.

7. Skills as the New Abstraction​

MCP "Skills" emerged as a new concept on the protocol roadmap — domain-specific knowledge bundled with MCP servers. Rather than just exposing tools, a skill packages the context, instructions, and workflow knowledge needed to use those tools effectively. This represents a shift from "here are the tools" to "here is how to accomplish this task." Skills are expected to become shareable primitives, enabling teams to distribute not just tools but expertise.

8. Evaluation-Driven Development​

Diamond Bishop (Datadog) delivered a clear message: you cannot improve what you do not measure. Datadog has deployed 100+ agents across products like Bits AI SRE, Bits AI Dev, and Security Analyst. Their approach treats evaluation as a living, breathing system — not a one-time gate — and makes eval infrastructure available via MCP so agents can contribute to their own improvement.

Other key lessons from Datadog: build agent-first interfaces (APIs and CLIs, not just UIs), favor proactive event-driven agents over reactive chat-based ones, use durable execution (Temporal) for production reliability, stay model-agnostic, and design for multiplayer (human-human, human-agent, agent-agent).

"Whoever is best today will probably not be best tomorrow. Build for portability." — Diamond Bishop, Datadog

9. The Agentic AI Foundation​

MCP governance has moved under the Agentic AI Foundation (AIF), which provides formal structure for the protocol's evolution. The foundation includes a Governing Board for strategic direction, a Technical Steering Committee for technical decisions, and 7 Working Groups with 500+ participants. The Identity & Trust Working Group alone has 150 members from 66 organizations. AIF has established a project lifecycle (Growth, Impact, Emeritus stages) for accepting new projects into the foundation.

10. Key Architectural Concepts​

Several architectural concepts appeared across multiple talks:

  • Gateway patterns: Access control, PII redaction, observability, and audit trails at the MCP interaction layer
  • Security layers: Defense in depth — MCP gateway, LLM gateway, supply chain filtering, isolated runtime, elevated access control
  • Agent runtime isolation: Enterprise agents should not run on developer desktops; they need sandboxed, controlled environments
  • Configuration as primitive: Agent configurations (which tools, which models, which permissions) treated as first-class, scannable, shareable objects
  • Durable execution: Temporal, containers, and sandboxes for production agent reliability

11. Company Highlights​

  • Anthropic: Co-creator of MCP. 100M+ SDK downloads/month. Protocol roadmap includes stateless transport, triggers, streaming, and skills.
  • AWS: Tens of thousands of internal builders using MCP. Internal registry for servers and agent configs. Core maintainer contributing to MCP spec. Introduced "lethal trifecta" security concept.
  • Uber: 5,000+ engineers, 1,500+ agents, 60,000+ executions/week. Built MCP gateway + registry. Three consumption surfaces: Agent Builder (no-code), Agent SDK (code-first), and Coding Agents.
  • Duolingo: 180+ tools, 30+ servers, 250+ weekly users on AI Slackbot. 80% approval rate. Open-sourced core at github.com/duolingo/slack-ai-agents.
  • Obot: CEO Sheng Liang (formerly Rancher Labs) positioned gateway + registry as essential enterprise infrastructure. Outlined five security layers.
  • Workato: Introduced Agent Control Plane concept and Seven Factors framework. White paper at workato.com/sevenfactors.
  • Datadog: 100+ agents in production. Bits AI SRE, Dev, and Security Analyst products. Emphasis on eval-driven development, model agnosticism, and durable execution.

The Road Ahead​

The conference pointed to several near-term developments:

  • Stateless transport for hyperscale deployments (1-2 months out)
  • Long-running tasks as a first-class protocol primitive
  • Cross-app authentication for seamless multi-service agent access
  • Triggers (webhooks) for proactive agent engagement
  • Streaming for incremental results
  • Skills as shareable, domain-specific knowledge packages
  • Formal governance through AIF working groups shaping the spec

The enterprise adoption curve is accelerating. Organizations that have not yet established their gateway + registry architecture and agent governance framework should expect to be doing so within the next 6-12 months.

Conclusion​

MCP Dev Summit NA 2026 made one thing unmistakable: MCP is the emerging standard for connecting AI agents to enterprise systems. The protocol is no longer experimental — it is in production at massive scale inside Amazon, Uber, Duolingo, and Datadog. The architectural consensus around gateways, registries, and control planes provides a clear blueprint for organizations beginning their MCP journey. With the Agentic AI Foundation providing governance and a robust roadmap of protocol improvements ahead, the ecosystem is positioned for rapid maturation over the coming year.

Conference: April 1-3, 2026 | New York City